Several fixes change results from earlier versions:
reli_stats(), reli_aov()) are now two-sided; a "95%" interval was previously 90%. Old results at level c equal new results at 1 - 2 * (1 - c).ccc_test(), agree_reps(), agree_nest()): new variance-components estimate for replicate data, and subject-level (jackknife) standard errors; the old intervals were too narrow.tolerance_limit()) with variance functions, clustered data or the bootstrap; prediction limits are unchanged unless a variance function is used.joint_test(), plot_joint(), and the joint power of deming_power_sim()/deming_sample_size()) now use Sadler's sum-of-squares F test. Version 0.3.0 used a chi-squared Wald test from the jackknife variance-covariance matrix, which rejected a true identity line up to 14% of the time at n = 10 (nominal 5%); method = "wald", test_method = "asymptotic" reproduces it.pb_reg(): error.ratio != 1 is now an error; bootstrap standard errors were 4-5 times too small before; tie-corrected slope intervals.find_n() can return a sample size one larger; agreement_limit(data_type = "reps") and "nest" change with unbalanced or missing data.icc_test() (one ICC as an htest, two-sided F intervals, REML for missing data, optional subject bootstrap), reli_err_test() (SEM, SEE, SEP and CVs), and mdc_test() (minimal detectable change).reli_facet_test() (experimental; its name may change) for subjects measured more than once by each rater or device: exchangeable repeats within subject-by-rater cells, or shared occasions (occasion, e.g. every rater on day 1 and day 2). Intra-rater, inter-rater and overall coefficients and SEMs from the generalizability-theory model, in consistency or agreement forms, with D-study projections, exact or MLS/Satterthwaite intervals, REML for unbalanced data, and the subject bootstrap. icc_test() and reli_err_test() point to it when the data have repeats, and mdc_test() accepts its SEMs.power_icc() and precision_icc() for ICC power and interval-width sample sizes, including missing data.summary(), plot() and check() methods for these results, with a new section in the power vignette.reli_stats() and reli_aov() are superseded and now use the same engine. Beyond the two-sided intervals: observed-data degrees of freedom with missing data, subject bootstrap, delta-method intervals for SEE/SEP/CV, ANOVA variance components, and stricter data checks.cac_test() (Gwet's AC1/AC2, Fleiss' kappa, Krippendorff's alpha, percent agreement; matches irrCAC) and ord_kappa_test() (model-based kappa for ordinal ratings from many raters, with group comparisons; #63), with summary() and plot() methods.agree_coef() is superseded by cac_test(); it gains categories and weights, and no longer fails or mislabels in edge cases.ccc_test(): concordance correlation coefficient as an htest, with an opt-in studentized subject bootstrap for replicate data.agreement_limit() gains bound_type ("iu" default, unchanged; "joint"), and prints within-subject SDs for replicate data.tolerance_limit():
bound_type, model = "lme" (including nested id), and conf_level;tol_method = "boot_cal", formerly "perc");model now works with update() and emmeans().loa_lme() is deprecated in favour of tolerance_limit(model = "lme").plot()/print() failing when settings were expressions (#50).power_agreement_limit() (Lu et al. 2016; reproduces PASS); blandPowerCurve() is superseded by it.power_agreement_exact(), agree_expected_half(), agree_assurance() and blandPowerCurve() gain m and icc for clustered data.n.joint_test() and plot_joint() gain method. For Deming fits the default is Sadler's (2010) sum-of-squares F test and its exact region. The jackknife Wald ellipse (method = "wald") rejected a true identity line about 8.5% of the time at n = 10 and 7.4% at n = 20 with the F reference (test_method = "F", its default), and 14% and 10% with the chi-squared reference that 0.3.0 used (nominal 5%); Sadler's test held 4.3-5.9% (references/general/eiv_sims). Passing-Bablok fits use the Wald test; Dufey (2020) fits are refused (8-11%).deming_power_sim() now uses joint_test() (new joint_method), so its power matches the test users run; it previously overstated joint power. deming_sample_size(use_joint = ) now takes effect.pb_reg():
se_method ("bootstrap", "jackknife", "dufey");error.ratio != 1 is now an error (it biased the slope in simulations; use dem_reg());id averages replicates and now works with missing values;reject_h0 on tied data.plot_joint() takes ideal_intercept before ideal_slope, like joint_test(), and both default to the model's confidence level.dem_reg(error.ratio) now correctly states var(x)/var(y), as computed (thanks to Hugo Lachuer).dem_reg function to include options for confidence intervals and plotting.pb_reg) for method comparison studies.simple_eiv objects to be more extensive and act like other statistical modelstolerance_limit and agreement_limitagreement_limit to allow of asymptotic confidence intervals to avoid errors related emmeans calculations
reli_stats documentation to remove references to BCa confidence intervalstolerance_limitagreement_limitagree_reps and agree_nest functions.vignette("reliability_analysis",package ="SimplyAgree") for details.chisq type of confidence intervals for "Other" CIs for reli_stats and reli_aovagree_np plot. (colors now align with order of LoA elements).reli_aov function for sums of squares approach.reli_stats with the cv_calc argument.
delta argument on plot if specified, and estimate of LoA.dem_reg) has been added to the package.agree_coef), otherwise known asagree_coef).loa_lme function
loa_mixed now deprecatedloa_lme allows for heterogenous variance by conditionagree_nest and agree_reps; now properly handles missing values